Matvos, Ostrovsky (2010) Mutual Fund Proxy Voting

Journal of Financial Economics 98 (2010) 90112
Contents lists available at ScienceDirect
Journal of Financial Economics

journal homepage: www.elsevier.com/locate/jfec
Heterogeneity and peer effects in mutual fund proxy voting$

Gregor Matvos a,, Michael Ostrovsky b
a b
University of Chicago Booth School of Business, Chicago, IL 60637, USA Graduate School of Business, Stanford University, Stanford, CA 94305, USA
a r t i c l e i n f o
Article history: Received 5 March 2009 Received in revised form 6 August 2009 Accepted 8 September 2009 Available online 1 April 2010 JEL classication: G34 D72 Keywords: Proxy voting Boards of directors Director elections Peer effects and strategic complementarities Supermodular games
abstract
This paper studies voting in corporate director elections. We construct a comprehensive data set of 2,058,788 mutual fund votes over a two-year period. We nd systematic heterogeneity in voting: some funds are consistently more management-friendly than others. We also establish the presence of peer effects: a fund is more likely to oppose management when other funds are more likely to oppose it, all else being equal. We estimate a voting model whose supermodular structure allows us to compute social multipliers due to peer effects. Heterogeneity and peer effects are as important in shaping voting outcomes as rm and director characteristics. & 2010 Elsevier B.V. All rights reserved.
The election of directors is the most important shareholder franchise. Larry Sonsini, Chairman, NYSE Proxy Working Group (NYSE, 2006).
1. Introduction Despite its importance, voting in the elections of corporate boards remains relatively unexplored. A major obstacle for detailed analysis of voting is the lack of
$ We thank Efraim Benmelech, Joshua Coval, Robin Greenwood, Guido Imbens, John Lazarev, Peter Reiss, and David Scharfstein, the anonymous referee, and especially Jeremy Stein, George Baker, and Malcolm Baker for helpful comments and suggestions. We are also grateful to Leonid Pekelis for excellent research assistance and to the Arthur and Toni Rembe Rock Center for Corporate Governance at Stanford University for nancial support. An earlier draft of this paper was circulated under the title Strategic proxy voting. Corresponding author. E-mail address: gregor.matvos@chicagobooth.edu (G. Matvos).
data on individual votes: until recently, voting was condential, with only the aggregate outcomes reported by the rms. Of course, aggregate data on voting outcomes can be very useful. For instance, Cai, Garner, and Walkling (2009) have used these data to study rm- and director-level determinants and consequences of director votes. It is, however, hard to gain insight into shareholderspecic determinants of voting using aggregate data. In 2003, the Securities and Exchange Commission (SEC) introduced a new rule, requiring mutual funds to report their votes. In this paper, we present the results of our analysis of mutual fund proxy voting behavior, based on the votes of the funds in the rst two years after the SEC rule change took effect. Our comprehensive data set contains 2,058,788 votes by 2,774 mutual funds in 13,588 director elections of 1,388 companies. This rich data set allows us to look at the behavior of individual voters. Our rst nding is that mutual funds systematically differ in their voting behavior. Some mutual funds are consistently more likely to cast votes in favor of directors sponsored by
0304-405X/$ - see front matter & 2010 Elsevier B.V. All rights reserved. doi:10.1016/j.jneco.2010.03.014
G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112
91
Table 1 Votes in director elections by 10 popular index funds. This table presents sample voting data for 10 popular mutual funds tracking the S&P 500 index in the elections of directors proposed by management. The votes are for the July 2003June 2004 and July 2004June 2005 voting periods. All votes other than for and withhold were discarded. Data source: SEC Edgar (N-PX lings). Mutual fund # For Vanguard 500 Index Fund USAA S&P 500 Index Fund Schwab S&P 500 Index Fund Merrill Lynch S&P 500 Index Fund Morgan Stanley S&P 500 Index Fund UBS S&P 500 Index Fund T. Rowe Price Equity Index 500 Fund Fidelity Spartan 500 Index Fund Smith Barney S&P 500 Index Fund Dreyfus S&P 500 Index Fund 2,686 2,992 2,791 3,200 3,183 2,954 2,942 3,089 2,920 3,176 July 1, 2003June 30, 2004 # Withhold 559 199 173 118 115 103 96 63 53 6 % Withhold 17.2% 6.2% 5.8% 3.6% 3.5% 3.4% 3.2% 2.0% 1.8% 0.2% # For 2,921 3,028 2,888 3,130 3,112 2,970 2,996 3,124 3,182 3,135 July 1, 2004June 30, 2005 # Withhold 351 223 208 107 130 80 112 38 42 15 % Withhold 10.7% 6.9% 6.7% 3.3% 4.0% 2.6% 3.6% 1.2% 1.3% 0.5%
the management than others. Our second nding is the presence of peer effects in mutual fund voting: a fund is more likely to oppose management when other funds are more likely to oppose it as well. These strategic interactions amplify funds equilibrium voting responses to factors that affect fund voting. For example, a negative change in director quality will rst have a direct effect: each fund is less likely to support a lower-quality director. But there is also an additional force: knowing that other funds are less likely to support the director, a fund has an additional reason to withhold its support. Thus, in equilibrium, the direct effect of any policy change will be magnied. It is instructive to contrast these ndings with hypothetical straightforward voting behavior. In the hypothetical case, all shareholders have the same incentive: to promote the behavior of directors that serves the best interests of the company. Each fund evaluates each director and then votes accordingly. Of course, even under such straightforward behavior, we would expect to see differences in fund voting behavior, simply due to random noise. However, we would not see systematic differences, and the identity or characteristics of the shareholders casting proxy votes would not play a role. In contrast, our results show that they matter. Moreover, the magnitudes of fund heterogeneity and peer effects are comparable economically to the effects of rm and director characteristics on voting outcomes. To show that some funds are systematically more management-friendly than others, we use a funds past voting record as an estimate of its friendliness. We nd that among funds who vote on the same director in the same meeting, funds with a higher estimate of friendliness are signicantly more likely to vote for. In other words, fund friendliness, as measured by the past voting record, is an important determinant of mutual fund voting in board of directors elections. A simple example of voting patterns provides an illustration of both heterogeneity and persistence of fund voting behavior. Table 1 presents the number of for (i.e., in support of a management-proposed director) and withhold (i.e., against a management-proposed director) votes for 10 large, popular mutual funds
tracking the S&P 500 index for two voting seasons. While the holdings of these funds are, by construction, very similar, the votes are not. The least managementfriendly fund, Vanguard 500 Index Fund, withheld support from management-proposed directors 559 times, or in 17.2% of cases, in the rst voting season, and 351 times, or in 10.7% of cases, in the second one. The corresponding numbers for the friendliest fund, Dreyfus S&P 500 Index Fund, are 6 (0.2%) and 15 (0.5%)lower than Vanguards by a factor of almost 100 in the rst voting season and almost 25 in the second. The 10 funds voting policies are also highly persistent: the correlation between their votes in the rst voting season and in the second one is equal to 0.93. There are several reasons why mutual funds may systematically differ in their management-friendliness. One is the degree to which they worry about potential management retaliation. In particular, funds may care about the current and potential future business ties with the rm, such as managing pension plans.1 Davis and Kim (2007) nd that fund families that derive a larger part of their revenue from management fees from their portfolio rms use voting policies that are friendlier to management. It can therefore be costly for a single fund to vote against directors recommended by management. However, management may have a hard time severing business relations with many funds holding its shares. Similarly, if the rm were to retaliate by other means, such as restricting funds access to their management,2 its ability to punish any individual fund would be diminished if it had to retaliate against a larger number of funds. In the extreme case, if all funds vote against directors
1 Other potential reasons include differences in proxy guidelines, differences in how proxy voting is organized and monitored within fund families, and fund manager individual differences in disutilities of opposing management and resisting the pressure to support management-nominated directors. 2 For example, in its comments to the SEC on vote disclosure rules, a mutual fund company states that this retaliation could be in the form of denial of access to company management in the course of our investment research on behalf of our shareholders. http://www.sec. gov/rules/proposed/s73602/rmason1.txt. Accessed July 27, 2009.
92
recommended by management, the management may resign and will not be able to retaliate. This safety in numbers externality can therefore be a source of peer effects in mutual fund proxy voting. After establishing persistent heterogeneity in fund voting behavior, we move on to testing for the presence of these peer effects and evaluating their magnitudes. Identifying peer effects requires a careful approach, because rm and director characteristics that are not observed by the researcher but are known to the funds induce a correlation in fund votes. This correlation can be mistaken for funds strategic interactions. We exploit the heterogeneity in funds management-friendliness to overcome this issue. Intuitively, suppose Fidelity is voting on two otherwise identical directors, except that the rst director will be voted on by management-friendly funds and the second director by unfriendly ones. The rst director will get more support from funds other than Fidelity than the second director, even though they are identical. Fidelity should not intrinsically care about how management-friendly other funds are; the managementfriendliness of the other funds is of interest only to the extent that it affects their votes. Therefore, if Fidelitys vote is more favorable for the rst director, this indicates that Fidelity raised its vote in response to the more favorable vote by other funds. This idea forms the basis of our estimation: we use management-friendliness of other funds in an election as an instrument for their expected vote. We rst estimate a basic linear instrumental variables model. We nd that funds are more likely to vote for when their expectations about the number of for votes cast by other funds are higher. To analyze the magnitudes of these effects in more detail, we also estimate a structural model of voting in board of directors elections, again exploiting the variation in fund friendliness to identify the parameters. The model incorporates our two main ndings from the reduced-form analysis: heterogeneity in fund friendliness and payoff complementarities (peer effects). We nd that the effect of mutual fund friendliness on voting in board of directors elections is of the same order magnitude as the effects of rm and director characteristics. This indicates that who votes on a director can potentially be as important as how good this director is. We also nd signicant peer effects in fund voting behavior. The following back-of-the-envelope calculation illustrates their magnitudes. Consider a fund that is likely to withhold its vote from management with probability 20%, taking into account that other funds will withhold their vote with probability 20% as well. Suppose that the funds expectation of how other funds will vote changes: it thinks other funds will withhold their vote with probability 10%, even though the funds own estimate of director quality remains the same. Our estimates suggest that in response to this change in expectations, the fund will withhold its vote with probability 14.9%. This same estimate also implies that the effects of changes in director quality will be amplied in equilibrium. Suppose a directors quality improves to such a degree that each fund in isolation would decrease
its probability of voting withhold from 20% to 10%. In addition, each fund now knows that all other funds will lower their vote by 10%. Given our estimates, in response to this change in other funds voting, each fund would further decrease its probability of a withhold vote from 10% to 7.2%. But then all funds know that as well, so they have an incentive to reduce their vote even further: an additional adjustment decreases the probability of funds voting withhold from 7.2% to 6.6%. Iterating this calculation eventually converges to each fund voting withhold with probability 6.4%. Thus, the 10% direct effect of changing director quality translates to a total effect of 13.6%: the social multiplier is 1.36, and the complete effect of changing a director characteristic is more than a third larger than it would be in the absence of peer effects. To evaluate the economic magnitude of peer effects in practice, we compute counterfactual equilibria using the estimated structural model. For each director from a subset, we rst compute the effect of an increase in quality if all funds voted in isolation. We then compute the equilibrium of the voting game for the same increase in director quality. We compare the two outcomes and nd that the multipliers are often large, and differ substantially across directors. For example, the median multiplier is about 1.07, but the 75th percentile multiplier can be as high as 2.13 (depending on the details of the model). This means that for many directors, the effect of changing director quality on voting outcomes more than doubles in the presence of peer effects. In addition to the results on proxy voting, our paper offers a methodological contribution to the literature on estimating static games with peer effects. We model voting in director elections as a supermodular game of asymmetric information (Van Zandt and Vives, 2007). This structure allows us to compute bounds on counterfactuals even in the presence of multiple equilibria. It also allows us to compute director-specic multiplier effects, taking into account director characteristics and shareholder composition. The computations can be performed using a simple, intuitive, and fast iterative process. Our paper is most closely related to two strands of corporate governance literature. The rst one is the literature on the determinants of board composition. These have been widely studied using outcome variables, such as board composition and director survival rates (see, e.g., Hermalin and Weisbach, 2003 for an excellent survey). Recent papers by Cai, Garner, and Walkling (2009) and Fischer, Gramlich, Miller, and White (2009) add to that literature by studying the effects and determinants of director elections using data on rmlevel voting outcomes. We contribute to this stream of research by looking at the individual votes cast by mutual funds, allowing us to examine persistent differences in funds behavior and the effects of interactions among funds on the voting outcome. The second strand of the literature empirically analyzes the effect of mutual fund incentives on their proxy voting behavior, using the data on individual votes that became available after the SEC rule change. Davis and Kim (2007) study voting on non-binding shareholder
93
proposals and nd that fund families that depend more heavily on business ties with portfolio rms use more management-friendly voting policies. Matvos and Ostrovsky (2008) show that mutual funds incentives to vote against bad mergers in acquiring companies are blunted because they realize a portion of merger gains in their holdings in the target, and that as a result rms with holdings in both the acquirer and the target are less likely to oppose such mergers than rms with holdings only in the acquirer. Rothberg and Lilien (2006) provide a variety of descriptive statistics about voting policies and outcomes on all voted issues for a sample of mutual fund families for the rst year after the SEC rule change took effect. We contribute to this strand of literature by focusing on director elections, constructing a comprehensive data set of fund votes in these elections, and showing that differences in voting patterns among funds are persistent and have a large impact on voting outcomes. Our paper also contributes to the literature on peer effects and social multipliers. These effects have been found in a variety of settings: crime (Glaeser, Sacerdote, and Scheinkman, 1996), education (Sacerdote, 2001; Cooley, 2010), stock market participation (Hong, Kubik, and Stein, 2004), mutual fund investment decisions (Hong, Kubik, and Stein, 2005), and many others. We show that these effects are important in corporate director elections. To the best of our knowledge, our two-step estimation approach, in which we rst establish persistent heterogeneity in funds behavior and then exploit that heterogeneity to identify peer effects, is novel. We also offer a convenient structural framework for analyzing them. The remainder of the paper is organized as follows. In Sections 2 and 3, we provide institutional background on mutual fund voting in board of directors elections and describe the data. In Section 4, we present reduced-form results demonstrating persistent differences in voting patterns among funds and the presence of peer effects. In Sections 5 and 6, we present a structural model of strategic proxy voting and discuss our estimation methodology for that model. We present our structural estimates in Section 7, and then in Section 8 use these estimates to conduct counterfactual simulations, helping evaluate the magnitudes of social multipliers due to peer effects. We conclude in Section 9.
proxy votes for its list of directors. Shareholders can prepare a competing list of directors by entering into a proxy ght, which they must nance out of their own pocket. Unless the corporation explicitly implements condential voting, the voting decisions of individual shareholders are revealed to management. If voting is condential, management is in principle informed only of the nal tally of the votes, and not the votes of individual shareholders. Condentiality, however, is not absolute and can be violated in a contested election.3 Also, after the 2003 policy change by the SEC, condentiality no longer applies to mutual funds, who are required to annually disclose their votes to the SEC in N-PX lings (Securities and Exchange Commission, 2003). These lings are subsequently made public by the SEC, and form the basis of our data set, described in Section 3. In the plurality voting system implemented by most corporations in the U.S., directors with the most for votes are elected. If directors run unopposed, shareholders cannot vote against a nominee, but can only withhold authority to cast the vote. The withhold vote therefore cannot prevent a nominees election, and even a single for vote theoretically elects an unopposed director.4 The withhold vote was introduced by the SEC and was later interpreted by it as a mechanism for shareholders to express their dissatisfaction with directors (Institutional Shareholder Services, 2005). While not legally binding, the withhold votes play an important role in disciplining management and allowing shareholders to publicly signal their dissatisfaction. Cai, Garner, and Walkling (2009) nd that poor election outcomes lead to reduced chief executive ofcer (CEO) compensation and increased CEO turnover, along with removal of anti-takeover mechanisms. Similarly, Fischer, Gramlich, Miller, and White (2009) show that lower voting outcomes lead to subsequent CEO and director turnover, along with changes in investment and acquisition policies of the rms. 3. Data Since 2004, the SEC has required mutual funds registered in the U.S. to annually report their votes in all shareholder meetings of their portfolio companies using N-PX lings. The deadline for reporting is August 31, and the votes cover the period beginning on July 1 of the previous year and ending on June 30 of the current year. For each vote, funds specify the company, the date of the meeting, a brief description of the issue being voted on (director election, merger proposal, shareholder proposal, etc.), the sponsor of the proposal (management or shareholders), managements recommendation, and the vote of
3 See Heard and Sherman (1987), McGurn (1989), and Monks and Minow (2003) for further details about the mechanics of proxy voting and an overview of vote condentiality issues. 4 This is not the case in the majority voting system, where withhold votes have the weight of an against vote and a candidate gets elected only if more than half of the votes cast are for (Institutional Shareholder Services, 2005).
2. Institutional background A board of directors in the United States is formally both a principal in its relationship to the management, and an agent with respect to the shareholders. The board is responsible for providing guidance and monitoring management on behalf of the shareholders. The nominees for boards of directors in the U.S. are selected for election by the board itself or by the boards nominating committee. The candidates, however, are often suggested by the companys management, which undertakes the effort to nd and evaluate them. Once the slate has been conrmed by the board, the company can start soliciting
94
the fund. We download all available N-PX lings for two 12-month periods: 7/1/20036/30/2004 and 7/1/2004 6/30/2005. We then use 34 computer scripts to extract data from the 100 largest active mutual funds, for several popular fund families, and any other mutual funds which used the same format. Next, we use an algorithm to extract the director votes. The details of these algorithms are available from the authors upon request. We obtain all stock price information from the Center for Research in Security Prices (CRSP) database. Industry benchmarks are calculated for all stocks with positive prices available on CRSP in the same two-digit Standard Industrial Classication (SIC) code. We use Compustat Industrial Annual les to construct accounting and nancial company information.5 To calculate Q and book-to-market ratios, we closely follow the variable construction used in Malmendier and Tate (2008). We obtain rm governance characteristics from the Investor Responsibility Research Center (IRRC) Governance database. We assign governance characteristics of a rm in 2003 and 2005 using data from 2002 and 2004, respectively. The voting data are matched with CRSP, Compustat, and IRRC data on CUSIPs and tickers. We obtain director information from the Board Analyst Directors database, which is matched using director last name, ticker, and year. We obtain Institutional Shareholder Services (ISS) recommendations from the ISS Voting Analytics database. Finally, we apply various cleaning procedures to remove duplicate and internally inconsistent records from the data. We restrict the sample to management-sponsored elections in which at least 10 funds cast votes and votes were cast using proxies from ordinary common shares (share codes 10 and 11). We are left with a sample of 2,058,788 director votes. 3.1. Summary statistics Our data set covers 13,588 director elections in 2,528 shareholder meetings of 1,388 companies. In these elections, we observe 2,774 mutual funds casting 2,058,788 votes. The summary statistics are presented in Panel A of Table 2. The average number of directors elected in a shareholder meeting is 5.38. The variation in the number of directors voted on in a shareholder meeting is substantial, ranging between two and 10 for the 10th percentile and the 90th percentile of elections, respectively. Each director is voted on by an average of 152 funds. A director election at the 10th percentile of the number of funds voting has 49 funds casting votes on a director, while the 90th percentile election has 300 funds casting their votes on a director. The average director in our sample receives support from 89.8% of funds. The distribution of voting outcomes,
5 For our book value of assets, we use total assets (item 6). We dene capex-to-assets as the ratio of capital expenditures (item 128) to total assets (item 6). We calculate cash ow-to-assets as income before extraordinary items (item 18) plus depreciation plus amortization (item 14) divided by total assets (item 6). We dene leverage as liabilities (item 181) over total assets (item 6) and return on assets (ROA) as income before extraordinary items (item 18) over total assets (item 6).
by director, is very skewed, as 96.5% of funds vote for the median director. The 10th percentile of directors, however, obtains only 59.5% support. The summary statistics of director characteristics in our sample are presented in Panel B of Table 2. Insiders (i.e., employees) represent 16.7% of directors up for election; 70.7% are outside directors, and 12.6% are outside related directors, whose employer has a nancial relationship with the rm or who are former employees of the rm.6 CEOs comprise 10.1% of our sample. Panel C shows the distributions of some characteristics of the rms in our sample. S&P 500 companies comprise 28.9% of our data set, but the sample spans a wide range of companies by size, protability, and other measures. Simple comparisons of means in Panel A of Table 3 provide the rst look at which director characteristics are correlated with voting outcomes.7 Two groups of directors are particularly likely to receive less support: founders and outside directors. Outside related directors receive only 83.01% of for votes (vs. 91.92% for insiders and 93.18% for unrelated outsiders). One possible explanation for the difference is that mutual funds understand that both outside related and inside directors are not free of conicts of interests, and therefore, they obtain a lower vote than unrelated directors. Moreover, outside related directors have available substitutes in outside directors, while inside directors, who have signicantly more information about the company, do not. Founders receive only 85.61% of for votes (vs. 91.68% for non-founders). Unlike other directors, they frequently own a signicant share of the company and, while on the board, can wield substantial control. Because they started the company, they may then be prone to promoting their own agenda, which may differ from maximizing the value of the company. While strong in a univariate setting, the founder effect becomes smaller in magnitude and statistically insignicant in our full specication in Section 7.2. CEOs and Chairmen are often the targets of shareholder discontent, but also have more power to retaliate against shareholders. CEOs get 1.64% more support, on average, than non-CEOs, while Chairmen get 1.14% more support than other directors. We should interpret the results on director characteristics, which are economically small but statistically signicant, with caution, since funds focus on various governance characteristics may change over time. For example, it is possible that when the media are focusing on executive compensation, funds pay more attention to votes on this dimension than they do at other times. We use only two years of voting data in our analysis, and so some of our results on director characteristics may be specic to that period and thus require future testing on a longer data set. Panels B and C of Table 3 present voting outcomes cut by rm characteristics. Firm size and S&P 500
6 For a small number of directors, our data do not contain information on whether they are insiders, outsiders, or related outsiders. 7 A more detailed analysis of these effects in the context of the estimated model of voting is presented in Section 7.2.
95
Table 2 Summary statistics. The sample in Panel A contains 2,528 board of director elections of 13,588 directors sponsored by the management between 2003 and 2005. The sample contains 2,774 mutual funds, with 6,136 fund-year observations. Number of directors up for election is the number of directors voted on in a shareholder meeting on a given date. Number of funds voting on a director is the number of funds casting votes on a director in a board of directors election on a given day. The average for vote per director is the percentage of funds casting a for vote in a given director election. Average for vote per fund-year is the percentage of for votes cast by a fund in a given year. Panel B contains data on 13,588 directors who were up for election to a board of directors between 2003 and 2005 and were recommended for election by the management. The sample in Panel C contains 2,528 rm-year observations on rms that held director elections between 2003 and 2005. The industry return is the value-weighted return of the rms two-digit SIC industry. Data sources: SEC Edgar (N-PX lings), CRSP, Compustat, IRRC governance database, Board Analyst directors database. Panel A: Election and fund characteristics Variable Number of directors up for election Number of funds voting on director Average for vote per director Average for vote per fund-year Panel B: Director-meeting characteristics Variable CEO Chairman Founder Inside director Outside director Outside related director Audit chair Audit member Compensation chair Compensation member Executive chair Executive member Governance chair Governance member Nominating chair Nominating member Panel C: Firm-year characteristics Variable Last year return ROA Assets Q Book-to-market Leverage Cash ow-to-assets Capex-to-assets S&P Mean 13.0% 4.0% 13,855 1.89 0.51 0.57 0.76 0.21 500 St. dev. 60.9% 10.1% 70,712 1.18 0.36 0.24 1.86 0.15 28.9% N 2,504 2,528 2,528 2,506 2,506 2,524 2,391 2,232 2,528 Mean 10.1% 9.8% 1.9% 16.7% 70.7% 12.6% 0.5% 36.9% 4.5% 35.8% 1.4% 15.3% 3.6% 30.2% 3.9% 33.9% N 13,588 13,588 13,588 13,176 13,176 13,176 13,588 13,588 13,588 13,588 13,588 13,588 13,588 13,588 13,588 13,588 Mean 5.38 151.52 89.8% 91.0% St. dev. 3.19 125.08 17.3% 14.7% p10 2 49 59.5% 80.3% Median 4 107 96.5% 95.1% p90 10 300 100.0% 100.0% N 2,528 13,588 13,588 6,136
membership reveal the largest differences. In the largest size quintile, 92.98% of the votes are for, while that number drops to 89.28% for the rms in the smallest size quintile. Similarly, directors in S&P 500 rms on average obtain 93.07% of for votes, while directors of other rms obtain 89.03%. There are several possible reasons why directors of larger rms obtain more votes. First, the standard free-rider problem may be more prominent in large rms, decreasing the benets to voting withhold. Second, larger rms may be more important clients for money managers, decreasing funds incentives to vote withhold. Third, the management of large rms might also have more funds at their disposal to hire proxy
solicitors and investor relations rms to help them manage the voting. Somewhat surprisingly, directors in better performing rms receive fewer for votes than directors in worse performing rms: 89.87% for the rms in the highest quintile of absolute returns versus 92.40% for the rms in the lowest quintile. Comparing rms on their returns relative to their industry paints a similar picture. The univariate results on performance are to a large degree reversed when we estimate the full specication in which we can properly interpret the magnitudes: we discuss those results in detail in Section 7.2.
96
Table 3 Average vote by director and rm characteristics. Panels A and B present the percentage of funds which cast for votes in director elections by director and rm characteristics. Column 1 presents the percentage of for votes cast by funds on directors without the characteristic, and column 2 the percentage of for votes cast on directors with the characteristic. The last column presents the difference in the means for the two groups. Panel C presents the percentage of votes cast for in director elections in the highest and lowest quintile of rms sorted on rm characteristics. The last column presents the difference in the mean of the highest and lowest quintile. Data sources: SEC Edgar (N-PX lings), CRSP, Compustat, IRRC governance database, Board Analyst directors database. (***) Indicates statistical signicance at the 1% level. Panel A: Director characteristics Statistic CEO Chairman Founder Audit member Compensation member Executive member Governance member Nominating member Inside director Outside director Outside related director Mean vote N Mean vote N Mean vote N Mean vote N Mean vote N Mean vote N Mean vote N Mean vote N Mean vote N Mean vote N Mean vote N for for for for for for for for for for for (1) No 91.42% 1,859,111 91.47% 1,865,170 91.68% 2,023,205 91.52% 1,309,106 92.15% 1,339,713 91.59% 1,713,586 91.72% 1,401,004 91.89% 1,354,773 (2) Yes 93.05% 199,677 92.60% 193,618 85.61% 35,583 91.67% 749,682 90.51% 719,075 91.51% 345,202 91.26% 657,784 90.96% 704,015 91.92% 325,724 93.18% 1,444,383 83.01% 241,373 (2)(1) Difference 1.64% (***) 1.14% (***) 6.07% (***) 0.15% (***) 1.64% (***) 0.08% 0.46% (***) 0.93% (***)
Panel B: Firm characteristics (1) No 89.03% 763,647 91.50% 1,661,176 (2) Yes 93.07% 1,295,141 91.93% 390,133 (2)(1) Difference 4.04% (***) 0.42% (***)
Statistic S&P 500 Return below industry quartile Mean vote for N Mean vote for N
Panel C: Firm characteristics (lowest versus highest quintile) (1) Lowest quintile 92.40% 355,779 89.28% 155,737 92.36% 511,885 (2) Highest quintile 89.87% 315,643 92.98% 886,864 92.73% 306,066 (2)(1) Difference
Statistic Return quintile Size quintile Book-to-market quintile Mean vote for N Mean vote for N Mean vote for N
2.53% (***) 3.71% (***) 0.37% (***)
4. Reduced-form results In this section, we rst show that funds differ in their management-friendliness, i.e., in how likely they are to vote for directors proposed by management, holding all else equal. These differences are persistent. We then exploit this persistent heterogeneity to establish the presence of peer effects in voting, using a linear instrumental variables approach. While this approach allows us to establish the presence of these effects, it is
not suitable for estimating and interpreting their magnitudes. In subsequent sections we address this issue by presenting and estimating a structural voting model.
4.1. Fund heterogeneity If funds votes responded only to director characteristics and behavior, we would not expect to observe
97
systematic differences in fund voting patterns.8 Of course, even if all funds had identical voting policies and preferences, but their decisions were subject to some random noise or idiosyncratic preferences, we would still observe some of them voting for the management more often than others, simply due to chance. In this section, we show that the differences in voting patterns are persistent. We construct the proxy for fund friendliness by using a funds past voting record. To proxy for a funds type in year t, we average its votes in year t 1. Furthermore, to avoid any mechanical correlation, we calculate the funds proxy by excluding its votes in the rm whose votes we are trying to explain. Our proxy is simple and is designed to avoid mechanical correlation. Despite its simplicity, this proxy performs well empirically. The past voting record of a fund in other companies is a highly statistically signicant predictor of fund voting. Column 1 in Table 4 shows the results from a basic regression of a funds vote on a director of company A on the funds average vote in the previous year on all directors in companies other than A.9 A 10% decrease in the proxy decreases the probability that a fund will vote for a director by 5.5%. From Table 2, the standard deviation of the distribution of the funds average votes is 14.7%. Hence, a two-standard-deviation increase in fund friendliness corresponds to an increase of 16% in the likelihood of the fund voting for a director. Of course, a funds past voting record could simply proxy for the characteristics of rms in which the fund invests and their directors. To control for this possibility, we include various rm- and director-level measures. Column 2 in Table 4 presents the results after including these controls: a funds vote is still correlated with its past voting record in other rms, and the magnitude of the coefcient is virtually unchanged.10 There is still a possibility that some directors are better than others, but that this quality is not captured by the controls in our data. To address this issue, we compare funds voting on the same director in an election, therefore controlling for all director characteristics. Funds whose average vote last year in other rms was higher
should be more likely to vote for this director. Columns 3 and 4 of Table 4 show that this is indeed the case. Column 3 reports the results with a separate xed effect for each shareholder meeting, and in column 4 there is a xed effect for each director election (usually, there are votes on multiple directors in each meeting). In all specications, the past voting record predicts fund voting in an election, and the effect is highly statistically signicant. In other words, some funds are more management-friendly than others, and their past voting records capture these differences. There are several possible reasons for this persistent heterogeneity in management-friendliness. First, business ties between fund families and portfolio rms change little over time (Davis and Kim, 2007). Second, the organization of voting and proxy guidelines in fund families remain relatively constant over time as well. Furthermore, a fund managers preferences should not change much over time, i.e., if a fund manager does not mind opposing managers one year, she will not experience a lot of disutility from opposing them the year after. Finally, a potential source of persistent heterogeneity is funds trying to build a reputation for managementfriendliness, which is useful only if it persists over time.
4.2. Peer effects Armed with a proxy for mutual fund managementfriendliness, we can now show that mutual funds are more likely to vote for a director if they think other funds are more likely to vote for her as well. We cannot simply regress a funds vote on the actual votes of other funds in the same election. Director characteristics observed by the funds are only partially known to the researcher, inducing correlation in votes from the researchers perspective even in the absence of strategic considerations. A similar problem is induced by unobservable rm characteristics, such as rms business ties to funds, or their ability to manage their relationships with investors. To circumvent these problems, we exploit the heterogeneity in funds management-friendliness. We estimate a reduced-form linear probability instrumental variables model: we are predicting a funds probability to vote for a director given the average vote of other funds voting on this director. Our instrument for the average vote of other funds is their estimated management-friendliness. The identication assumption in this approach is that funds care about other funds management-friendliness only to the extent that it inuences other funds votes. Because more management-friendly funds are more likely to vote for a director holding all else equal, we can exploit this variation to generate variation in other funds expected votes. If a fund is to vote in two identical rms, where one is held by friendly funds and the other by unfriendly funds, it would vote for more frequently in the former if there were strategic considerations present. We discuss the validity of the identication assumption in Section 4.3. We begin with the most basic specication of the regression (Table 5, column 1). In that specication, the
8 Preferences of mutual funds may differ in mergers and acquisitions: e.g., two funds may have the same stakes in the acquirer and different stakes in the target, resulting in different preferences over the outcome of the proposed acquisition (Matvos and Ostrovsky, 2008). However, such differences in preferences seem unlikely for director elections. 9 The number of observations in this regression is 1,766,982, which is less than the total number of votes in our data set, 2,058,788. This is because the only votes on the left-hand side of this regression are the ones for which we can estimate the voting funds managementfriendliness based on the prior year. 10 We cannot interpret the effects of the controls on voting in this specication. First, the regression specication does not take into account how peer effects affect voting, but subsumes them with director xed effects. Second, the underlying non-linearity of voting is exacerbated by the fact that the mean probability of voting for is very close to one, making the linear approximation problematic. Because of nonlinearity, we also do not interpret the effects of controls in Section 4.2. We defer this discussion to Section 7.2, where we estimate the coefcients in a structural model and can interpret their magnitudes.
98
Table 4 Past vote as a measure of management-friendliness. The sample consists of 1,766,982 votes by mutual funds in director elections in years 2004 and 2005. The dependent variable is a dummy that takes the value of one if the vote is for and zero otherwise. Fund management-friendliness is the average vote of the fund in the previous year in director elections in rms other than the one under consideration. The governance index is the governance index from Gompers, Ishii, and Metrick (2003). The omitted category from Outside director and Outside related director is an Inside director. Standard errors are clustered on fund-year in columns 1 and 2, on the shareholder meeting in column 3, and on director election in column 4. Data sources: SEC Edgar (N-PX lings), CRSP, Compustat, IRRC governance database, Board Analyst directors database. (1) Vote for Fund management-friendliness Last year return ROA Industry return Return below industry quartile Log assets Q Book-to-market Leverage Cash ow-to-assets Capex-to-assets S&P 500 Governance index Outside director Outside related director CEO Chairman Founder Audit chair Audit member Compensation chair Compensation member Executive chair Executive member Governance chair Governance member Nominating chair Nominating member Constant Observations R2 Shareholder meeting xed effects Director election xed effects Robust standard errors in brackets. 0.419 [0.0381] 1,766,982 0.08 0.554 [0.0410] (2) Vote for 0.551 [0.0417] 0.006 [0.0014] 0.210 [0.0361] 0.001 [0.0053] 0.002 [0.0013] 0.001 [0.0007] 0.003 [0.0005] 0.007 [0.0026] 0.037 [0.0030] 0.131 [0.0351] 0.082 [0.0176] 0.029 [0.0027] 0.001 [0.0002] 0.055 [0.0020] 0.054 [0 0018] 0.022 [0.0015] 0.016 [0.0008] 0.023 [0.0022] 0.000 [0.0000] 0.014 [0.0006] 0.003 [0.0019] 0.032 [0.0020] 0.007 [0.0021] 0.010 [0.0008] 0.020 [0.0049] 0.003 [0.0017] 0.022 [0.0051] 0.018 [0.0017] 0.379 [0.0389] 1,493,621 0.11 (3) Vote for 0.543 [0.0118] (4) Vote for 0.542 [0.0043]
0.043 [0.0062] 0.057 [0 0075] 0.014 [0.0054] 0.009 [0.0042] 0.004 [0.0137] 0.000 [0.0000] 0.009 [0.0031] 0.005 [0.0056] 0.027 [0.0034] 0.006 [0.0065] 0.000 [0.0047] 0.004 [0.0253] 0.001 [0.0079] 0.006 [0.0250] 0.020 [0.0079] 0.424 [0.0122] 1,726,148 0.11 Y
0.429 [0.0039] 1,766,982 0.11 Y
Signicant at 10%; signicant at 5%; signicant at 1%.
99
Table 5 Determinants of voting behavior: reduced-form instrumental variables regression. The dependent variable is a vote cast by a fund in a board of directors election; it takes the value of one if the vote is for and zero otherwise. The vote of other funds is the average vote of other funds voting on the director. The instrument for the vote of other funds is: the average management friendliness of other funds in columns 1, 2, and 5; and the mean, standard deviation, skewness and kurtosis of the management-friendliness of other funds in columns 3, 4, and 6. The management-friendliness of other funds is calculated as the average vote of a fund in the previous calendar year on rms other than the rm under observation. The industry return is the value-weighted return of the rms two-digit SIC industry. The governance index is the governance index from Gompers, Ishii, and Metrick (2003). The ISS Recommendation is one if Institutional Shareholder Services recommended a for vote on a director and zero otherwise. The omitted category from Inside director and Outside related director is an Outside director. Data sources: SEC Edgar (N-PX lings), CRSP, Compustat, IRRC governance database, Board Analyst directors database. (1) Vote for Vote of other funds Firm characteristics Last year return Industry return ROA Return below industry quartile Log assets Q Book-to-market Leverage Cash ow-to-assets Capex-to-assets S&P 500 Governance index Director characteristics Inside director Outside related director CEO Chairman Founder Audit member Compensation chair Compensation member Executive chair Executive member Governance chair Governance member Nominating chair Nominating member ISS Recommendation Fund-year xed effects Observations Y 1,766,982 Y 1,493,620 Y 1,766,982 Y 1,493,620 0.6324 [0.0331] (2) Vote for 0.3246 [0.0386] 0.0036 [0.0012] 0.0121 [0.0036] 0.1145 [0.0340] 0 [0.0012] 0.0012 [0.0005] 0.0009 [0.0004] 0.0028 [0.0021] 0.0241 [0.0030] 0.0601 [0.0314] 0.0493 [0.0166] 0.025 [0.0023] 0.0004 [0.0002] 0.0369 [0.0024] 0.074 [0.0041] 0.015 [0.0013] 0.0105 [0.0009] 0.0137 [0.0022] 0.0091 [0.0007] 0.0004 [0.0019] 0.0217 [0.0022] 0.004 [0.0016] 0.0065 [0.0007] 0.0141 [0.0045] 0.0021 [0.0015] 0.0172 [0.0047] 0.0118 [0.0016] (3) Vote for 0.7041 [0.0316] (4) Vote for 0.5857 [0.0336] 0.0016 [0.0012] 0.0097 [0.0034] 0.071 [0.0325] 0.0005 [0.0012] 0.0011 [0.0005] 0.0002 [0.0004] 0.0012 [0.0020] 0.0145 [0.0028] 0.0337 [0.0301] 0.0299 [0.0159] 0.0132 [0.0021] 0.0002 [0.0002] 0.0228 [0.0021] 0.0457 [0.0034] 0.0095 [0.0013] 0.0062 [0.0008] 0.0076 [0.0020] 0.0056 [0.0006] 0.0004 [0.0019] 0.0137 [0.0021] 0.0029 [0.0016] 0.0043 [0.0007] 0.0087 [0.0043] 0.0016 [0.0015] 0.0106 [0.0045] 0.0076 [0.0015] (5) Vote for 0.0947 [0.0486] 0.0014 [0.0011] 0.0039 [0.0033] 0.0865 [0.0319] 0.0012 [0.0012] 0.0018 [0.0005] 0.0011 [0.0004] 0.0022 [0.0020] 0.0065 [0.0025] 0.0314 [0.0296] 0.0274 [0.0158] 0.0052 [0.0017] 0.0004 [0.0002] 0.0103 [0.0015] 0.0124 [0.0016] 0.0033 [0.0011] 0.0019 [0.0007] 0.0014 [0.0018] 0.0033 [0.0005] 0.001 [0.0019] 0.0114 [0.0021] 0.0019 [0.0016] 0.0054 [0.0006] 0.0106 [0.0042] 0.0004 [0.0014] 0.0068 [0.0043] 0.0009 [0.0014] 0.4718 [0.0261] Y 1,493,112 (6) Vote for 0.1627 [0.0469] 0.0014 [0.0011] 0.0031 [0.0033] 0.0801 [0.0318] 0.001 [0.0012] 0.0017 [0.0005] 0.0011 [0.0004] 0.0019 [0.0020] 0.0059 [0.0025] 0.0284 [0.0295] 0.0253 [0.0157] 0.0044 [0.0017] 0.0003 [0.0002] 0.0096 [0.0015] 0.0115 [0.0016] 0.0031 [0.0011] 0.0017 [0.0007] 0.0015 [0.0018] 0.003 [0.0005] 0.0009 [0.0019] 0.0106 [0.0021] 0.0016 [0.0016] 0.0051 [0.0006] 0.0098 [0.0042] 0.0004 [0.0014] 0.0063 [0.0043] 0.0009 [0.0014] 0.4364 [0.0254] Y 1,493,112
Robust standard errors clustered on fund-year in brackets. Signicant at 10%; signicant at 5%; signicant at 1%.
100
instrument for the average expected vote of other funds in an election is their average friendliness, i.e., the average of their average votes in the previous year in other rms. We regress a funds vote on a particular director on the instrumented average vote of other funds, controlling for fund-year xed effects. The coefcient on the average vote of other funds is positive at 0.63 and statistically signicant at 1%. This result implies that if the average vote of other funds increases by 10%, a fund becomes 6.3% more likely to vote for a director. In column 2, we add several rm and director characteristics that might inuence a funds vote: rm performance characteristics, standard nancial controls, and information on the director, such as her position in the rm, whether she is an insider, and which committees she is a member of. After adding these controls, the coefcient decreases to 0.32, but is still statistically signicant at 1%. Next, we consider an alternative way of instrumenting for the average vote of other funds. Managementfriendliness may impact future votes non-linearly, and while the average friendliness of other funds may predict their average vote well, other moments of the distribution of their friendliness may contain additional information. In the alternative specication, in the rst stage, we add three additional moments of the distribution of other funds management-friendliness as instruments: the standard deviation, skewness, and kurtosis.11 The results from this alternative specication are presented in columns 3 and 4. The estimated coefcient on the average vote of other funds is higher in these two regressions than in the corresponding regressions that only use the average management-friendliness as an instrument, and are highly statistically signicant. Our nal set of specications includes an additional control: voting recommendations of Institutional Shareholder Services (ISS), a popular proxy voting advisor.12 Including their recommendation may control for a potentially important part of the mutual funds information about the quality of directors. This variable, however, has an important drawback for our purposes: one of the key inputs that ISS uses in forming their recommendations are discussions with the institutional shareholders of that directors company.13 Therefore, the interaction between fund votes and ISS recommendations occurs in both directions. First, ISS recommendations may reect director quality and provide information to the funds, inuencing their votes. At the same time, ISS recommendations may incorporate the expected vote of other funds, thereby acting as a proxy for other funds vote. It is therefore hard to interpret regression coefcients on this variable and the impact of its inclusion on the coefcients on other variables, especially given the non-linear
structure of the environment. While we cannot disentangle these two effects, including ISS recommendations in the regressions is a useful robustness check, and we perform it in columns 5 and 6 of Table 5. In these specications, coefcients on the average vote of other funds are smaller in magnitude (0.09 and 0.16) than those in columns 2 and 4, but both of them are still statistically signicant, conrming the presence of peer effects. 4.3. Alternative explanations Before proceeding to our structural model, we rst address some alternative explanations of the voting complementarity discussed in the previous section and show that they are unlikely to hold. Our main concern is that mutual funds do not randomly select stocks for investment. If friendly funds hold different stocks from unfriendly ones, then holdings could induce a correlation in fund voting that would be mistaken for strategic interaction. In particular, if unfriendly funds hold companies in which directors have a low unobservable quality, our identication assumption would be invalid: A rm held by friendly funds would have better directors than whats predicted by its observable characteristics, and so the positive coefcient on the average vote of other funds would arise simply because the friendliness of other funds (and hence, their predicted average vote) would proxy for unobserved rm or director characteristics. Clearly, stock selection cannot be the only explanation of heterogeneous voting behavior by the funds: as pointed out in Section 1, even S&P 500 index funds, which have little choice regarding their investments, exhibit a high degree of heterogeneity in their average votes. However, to show that it is at best a minor driver of voting behavior and is unlikely to explain the strategic effects discussed in the previous section, we need a more detailed analysis. First, we examine which observable rm characteristics are correlated with the average friendliness of funds holding the rm. If unfriendly funds are more likely to acquire shares in lower-quality companies, we should expect them to hold rms that are worse on both observable and unobservable characteristics. In fact, observable rm characteristics such as past returns and governance characteristics should be particularly strongly correlated with fund friendliness, if we expect unobservable rm quality, which is neither captured by accounting numbers nor by stock returns, to matter as well. Table 6 presents the correlation of observable rm characteristics with the average friendliness of funds holding shares in the rm. If unfriendly mutual funds acquire shares in lower-quality companies, one would expect these rms to also exhibit bad past performance. This is not the case. Neither last years stock return nor the rms accounting return on assets is positively correlated with the average management-friendliness of funds holding shares in the rm. Similarly, looking at relative returns, rms in the bottom quartile of their industries are not more likely to be held by unfriendly funds. If anything, the coefcients in these regressions (columns 2 and 4) suggest that worse performing rms are held by friendlier
11 We have also considered other parameterizations of the distribution. Details are available upon request. 12 In 2007, ISS was acquired by RiskMetrics Group. 13 For instance, on its Web site, RiskMetrics advertises engagement with appropriate company ofcials, top institutional holders and other parties to gain insight and make informed vote recommendations. http://www.riskmetrics.com/proxy_advisory/benets. Accessed July 27, 2009.
101
Table 6 Average management-friendliness of funds holding a rm. The dependent variable is the average management-friendliness of mutual funds holding shares in a rm. A funds management-friendliness is calculated as the average vote of a fund in the previous calendar year on rms other than the rm under observation. The industry return is the valueweighted return of the rms two-digit SIC industry. The governance index is the governance index from Gompers, Ishii, and Metrick (2003). The omitted category from Outside director and Outside related director is an Inside director. Data sources: SEC Edgar (N-PX lings), CRSP, Compustat, IRRC governance database, Board Analyst directors database. (1) Average managementfriendliness Last year return ROA Industry return Return below industry quartile Log assets Q Book-to-market Leverage Cash ow to assets Capex-to-assets S&P 500 Governance index Constant Observations R2 0.868 [0.003] 2,348 0.21 0.867 [0.003] 2,324 0.34 0.02 [0.001] 0.021 [0.001] 0.02 [0.001] 0 [0.000] 0.866 [0.004] 2,185 0.22 0.003 [0.000] 0.003 [0.000] 0.003 [0.000] 0.041 [0.002] (2) Average managementfriendliness 0.002 [0.001] (3) Average managementfriendliness (4) Average management friendliness 0.002 [0.001] 0.014 [0.022] 0.038 [0.002] 0.002 [0.001] 0.005 [0.000] 0.001 [0.001] 0.004 [0.002] 0.007 [0.003] 0 015 [0.023] 0.009 [0.015] 0.016 [0.001] 0 [0.000] 0.859 [0.004] 1,926 0.38
Robust standard errors clustered on fund-year in brackets.
funds, although none of them are statistically signicant in the full specication. Combined, these results do not support the view that unfriendly funds hold shares in bad companies. Instead, mutual fund friendliness is correlated with the book size of a rms assets, its leverage, book-tomarket, membership in the S&P 500, and the industry return. These variables are highly correlated with the determinants of mutual funds management styles as in Goetzmann and Brown (1997). In other words, the correlation of average mutual fund friendliness and rms observable characteristics is consistent with the standard determinants of fund style, but not with the hypothesis that unfriendly funds purchase lower-quality companies. Another way to check whether fund friendliness is correlated with unobservable director quality in the funds portfolio companies is the following. As we mentioned above, if unfriendly funds hold companies that are bad on an unobservable characteristic, then a funds past voting record is not only proxying for the funds friendliness, but also for the rms unobservable quality. Suppose that is indeed the case. Then, to predict the quality of a rm, one has to look at both its observable
characteristics and at which funds hold the rm. If the funds holding the rm frequently voted against other rms in their portfolio in the past, then this rm is also a part of a portfolio of bad rms and is therefore, on average, worse than one would predict from its observable characteristics. A funds past voting record in other rms then contains information both about the funds actual friendliness and about the quality of the rm in question. When we predict a funds voting behavior using its past voting record, as we did in Table 4, column 2, controlling for observable rm and director characteristics, the funds past voting record is potentially correlated with its current vote for two different reasons. First, a high past voting record in other rms is correlated with a funds vote because the fund is management-friendly. Second, the past voting record in other rms is correlated with a funds vote in the rm under consideration because all rms in the portfolio have similar unobserved quality. Now, if instead we run an alternative specication with an election xed effect, that xed effect should absorb all rm information that was up to this point contained in a funds past voting record. Hence, if a funds past voting record in other companies were proxying for unobserved
102
rm quality, then its effect on a funds vote should be diminished after the inclusion of the xed effect. Therefore, one would expect the coefcient on the past voting record to disappear or at least to decrease substantially after the inclusion of director xed effects. When we run this alternative specication, the size of the coefcient is barely affected (column 4 of Table 4). The coefcient changes from 0.551 in the specication with rm and director characteristics (column 2), to 0.542 with director election xed effects (column 4). The standard error of the noisier coefcient is 0.04, and so the change of 0.009 is statistically indistinguishable from zero. Thus, all our evidence points to the strategic driver of the effect presented in the previous subsection, with funds incorporating their expectations about other funds votes into their own decisions, rather than the alternative explanation that the past voting record of a fund in other companies proxies for an unobserved quality of the rm in question.
each fund realizes its payoff U qi , nj , oij , oi, j , eij : 1
The payoff of the fund depends on director characteristics, the funds management-friendliness, its own vote, the votes of other funds, and the funds director-specic shock. Note that we assume that only the funds own friendliness enters the payoff function of a fund directly. The fund cares about other funds management-friendliness only to the extent that it affects other funds votes. This assumption formalizes our exclusion restriction behind the reduced-form estimates. We also make several additional assumptions, discussed below. Assumption 1. Voting has increasing differences in director characteristics, i.e., for all k, @U qi , nj , 1, oi, j , eij U qi , nj , 0, oi, j , eij Z 0: @qik 2
5. The model of strategic proxy voting We now turn to the structural analysis of strategic proxy voting. We model voting as a simultaneous-move game.14 There are ni funds, j 2 f1, . . . , ni g, voting in electing director i to the board.15 The director has a vector of characteristics qi qi1 , . . . , qim , representing both director- and rm-specic characteristics. All characteristics are observed by all funds. Each fund has a certain management friendliness parameter nj . As mentioned above, funds differ in their approaches to voting. Friendliness nj captures such systematic differences on a scale of how favorable these are to management. Formally, funds with a higher nj , holding all else equal, have a higher relative payoff to voting for vs. withhold than less friendly funds. The funds friendliness nj is known to all funds. The fund also privately observes an idiosyncratic director-specic shock eij , which affects its payoff from voting.16 This shock could arise, for example, if a fund manager nds it particularly costly to oppose a particular director due to a personal or business connection. These shocks are drawn independently from a zero-mean distribution Geij . Funds simultaneously cast their votes on director i. We denote by oij 2 f0, 1g the vote of fund j on director i, with 1 representing a for vote and 0 a withhold vote. The vector of all votes cast for director i is then oi oi1 , . . . , oin ; also, let oi, j oi1 , . . . , oi, j1 , oi, j 1 , . . . , oin . After votes are cast,
14 We thus rule out the possibility that a fund observes the votes of others in a given election before making its own decision. We also, in effect, view each election in isolation, assuming away dynamic strategies that funds may play in the repeated game of voting in multiple elections over time. 15 For notational convenience, we will sometimes omit index i. 16 We interpret es as only inuencing the private values of the funds from voting and not as private signals about underlying director qualities. The latter interpretation would imply a voting game with information transmission and interdependent valuations, similar to the model of Feddersen and Pesendorfer (1997).
That is, we assume that higher values of characteristics qik are viewed positively by the funds, and that all else being equal, the fund is more willing to vote for a director with higher values of these characteristics. Assumption 2. Voting has increasing differences in management-friendliness and the idiosyncratic shock, i.e., @U qi , nj , 1, oi, j , eij U qi , nj , 0, oi, j , eij Z 0, @nj @U qi , nj , 1, oi, j , eij U qi , nj , 0, oi, j , eij Z 0: @eij 3
This assumption formalizes the notion of the friendliness parameter and the role of the idiosyncratic shock: a more friendly fund has a higher incremental payoff of voting for, holding director characteristics and votes of other funds xed. It also states that a higher idiosyncratic shock increases the incremental return to voting with management. Assumption 3. Voting has increasing differences in other funds votes, i.e., U qi , nj , 1, oi, j , eij U qi , nj , 0, oi, j , eij Z U qi , nj , 1, o0i, j , eij U qi , nj , 0, o0i, j , eij whenever, vote by vote, oi, j Z o0i, j . This assumption introduces a specic form of inter-fund externalities: voting withhold is more costly when few other funds also vote withhold, for the reasons outlined in Section 1. The presence of these externalities, in turn, implies that in the voting game, funds will behave strategically: all else being equal, a fund is less likely to vote against a director if it expects few other funds to vote withhold. While this voting game can have multiple equilibria, Assumptions 13 above imply that they have a special structure: there exists a pure-strategy equilibrium most friendly to management, and a pure-strategy equilibrium most hostile to management. All other equilibria are contained between these two extremes. Proposition 1 states this result formally: 5
103
Proposition 1. There exist the lowest and the highest purestrategy Bayesian Nash equilibria of the voting game specied above. (These equilibria may coincide.) In both equilibria, each funds vote is weakly increasing in the funds own friendliness parameter nj , other funds friendliness parameters nk , kaj, and the vector of director qualities qi . Proof. The game satises the assumptions of a monotone supermodular game in Van Zandt and Vives (2007):
Condition 1 requires supermodularity and increasing

differences in actions and parameters, which determine a funds payoff qi1 , . . . , qim , nj , eij , which are satised by Assumptions 13. Condition 2 requires that the beliefs of fund j about vectors qi1 , . . . , qim , nk , eik for all kaj are rst-order stochastically increasing given its own realization of qi1 , . . . , qim , nj , eij . Given that the idiosyncratic shocks eij are independent, management-friendliness of funds n is commonly known, and the characteristics qi are common to all funds, this condition is trivially satised.
reduced-form model of Section 4.2. We circumvent it using the same identication strategy: exploiting the heterogeneity in funds management-friendliness. If our best response function (6) were linear, our reduced-form approach from Section 4.2 would estimate its structural parameters. However, because our outcome variable is binary, our structural equation is inherently non-linear. Hence, the linear instrumental variables approach can only provide a rough approximation of the best response function. For our main estimation procedure, we use an approach based on the control function methodology (Imbens and Wooldridge, 2007; Navarro, 2008). We start by making several specic functionalform assumptions about the funds payoff function U qi , nj , oij , oi, j , eij . First, we assume that the votes of other funds enter the utility function in a particular way: a fund cares about the fraction of other funds voting for, i.e., since each vote can take the value of either one or zero, each fund cares about the average vote of other P funds, o i, j 1=ni 1 kaj oik , where ni is the total number of funds voting on director i. In other words, U qi , nj , oij , oi, j , eij U qi , nj , oij , o i, j , eij : 7
Proposition 1 now follows directly from Theorem 1 in Van Zandt and Vives (2007). & Remark 1. In equilibrium, funds play threshold strategies in qi , nj , and eij . Remark 2. The best response function of a fund is increasing in qi , nj , eij , and the expected vote of other funds. Generally, best response functions in games are specied in terms of strategies of other players, which are not directly observable. Remark 1, however, allows us to express a funds strategy as a cutoff strategy in eij , conditional on commonly known director quality qi and the funds friendliness nj . Therefore, there is a monotonic one-to-one relationship between a funds strategy and its expected vote, conditional on qi and nj , and we can replace other funds strategies with their expected votes when writing down a funds best response function. 6. Estimation We now turn to estimating fund voting behavior, i.e., in the framework of Section 5, a funds best response function:
o ij oij qi , nj , Eei, j oi, j , eij : ij
Second, we divide the vector of director characteristics qi into a vector of characteristics observable by both the researcher and the funds, xi, and a scalar representing unobserved director quality, zi , which is known to the funds, but not to the econometrician: U U xi , zi , nj , oij , o i, j , eij : 8
Third, we normalize a funds utility from voting withhold to zero: U xi , zi , nj , 0, o i, j , eij 0: 9
Finally, we assume that the utility from voting for is linear: U xi , zi , nj , 1, o i, j , eij a b1 nj b2 o i, j b3 zi Gxi eij : 10 This assumption precludes non-linear interactions between various parameters of interest. It is, however, a common restriction in the literature on estimating discrete choice models and discrete games, which makes the estimation problem computationally tractable (see e.g., Ackerberg, Benkard, Berry, and Pakes, 2007). Also, in an earlier version of the paper, we have estimated several more exible specications, in which we included quadratic and cubic terms of parameters and their interactions. Our results are robust to those specications and, if anything, are stronger economically and statistically; details are available upon request. When fund j casts its vote, it does not know how the other funds will vote. Those funds votes depend on the realizations of their idiosyncratic shocks eik , kaj. Let ei, j denote the random vector of these idiosyncratic shocks. Fund j maximizes its expected utility Eei, j U xi , zi , nj , oi, j , o i, j , eij . If it votes withhold, its utility is zero: Eei, j U xi , zi , nj , 0, o i, j , eij 0: 11
Best response o is an increasing function of director characteristics, fund characteristics, other funds expected votes, and idiosyncratic fund-director shocks (see Remark 2). Ideally, to estimate the function, one would observe all variables that enter it. Then one would directly estimate all the parameters of the model, including the parameter on other funds expected vote, to test for the presence of strategic interactions. The problem we encounter is that director characteristics observed by the funds, qi, are only partially known to the researcher. This is the same issue as we encounter when we estimate peer effects in the
104
If instead it votes for, its expected utility is equal to Eei, j U xi , zi , nj , 1, o i, j , eij Eei,j a b1 nj b2 o i, j b3 zi Gxi eij a b1 nj b2 Eei, j o i, j b3 zi Gxi eij : 12 Let o e i, j denote the average expected vote of other funds, Eei, j o i, j . Note that this expectation is conditional on the information of fund j, i.e., funds friendliness parameters n and director is characteristics xi and zi . Eqs. (11) and (12) imply that the best response function of fund j, o i, j , takes the form ( ) 1 if a b1 nj b2 o e i, j b3 zi Gxi eij Z 0 o : 13 e ij 0 if a b1 nj b2 o i, j b3 zi Gxi eij o 0 The vote of fund j on director i, o is determined by the funds own management friendliness nj , other funds expected average vote o e i, j , director is observed characteristics xi and unobserved quality zi , and fund-director random shock eij . Variables zi (unobserved director quality) and o e i, j (expected vote of other funds) are not observed by the researcher, and so we cannot estimate the coefcients in Eq. (13) directly. Instead, we rst estimate these two variables, and then in the second stage estimate the parameters of interest. To separately identify these two variables, we use instruments that affect o e i, j but not zi . The key assumption behind our estimation approach is that other funds friendliness parameters do not directly enter a funds best response function (13). Therefore, we can use management-friendliness parameters of funds other than j, ni, j , to instrument for their votes. We can then use variation in ni, j to identify the strategic interaction parameter b2 and other coefcients in Eq. (13).
ij ,
The actual, observed average vote of other funds is equal to the expected vote plus noise (due to idiosyncratic shocks in the preferences of other funds),
e o i, j o i, j Zij :
15
Unlike the funds, the researcher does not know the unobserved component of director quality, zi , and hence, from his point of view, the expected vote of other funds is equal to
or i, j d g1 nij g2 n i, j G1 xi :
16
We can now rewrite the actual average vote of other funds as

e o i, j o i, j Zij
17 18 19 20
d g1 nij g2 n i, j G1 xi g3 zi Zij d g1 nij g2 n i, j G1 xi g3 zi Zij or i, j rij ,
where rij g3 zi Zij is orthogonal to n and xi by construction. Hence, the researchers expectation o r i, j and parameters d, g1 , g2 , and G1 can be consistently estimated by regressing the realized average vote of other funds, o i, j , on n and xi . The estimate of the researchers expectation ^ g ^ r d ^ 1 nij that we obtain from this regression is o i, j ^ 1 xi . ^ 2 n i , j G g Next, we can also obtain an estimate of unobserved director quality. Namely, we can rewrite the above equations as
r g3 zi o i, j o i, j Zij ,
21
6.1. First stage Fix the equilibrium. The average expected vote of other funds, o e i, j , is then a function of funds friendliness parameters, n, and director is observable and unobservable e 17 In characteristics xi and zi , i.e., o e i, j o i, j n1 , . . . , nn , xi , zi . principle, we could have derived the shape of this function for various values of parameters from equilibrium considerations and then applied a moment- or likelihood-based procedure to estimate the parameters of interest by maximizing the likelihood or minimizing the deviations from moment conditions of the observed voting behavior. This, however, would be very demanding computationally, and so we instead use several alternative parametric approximations of function o e i, j and subsequently verify that our results are robust to the choice of the functional form of the approximation. In the basic approximation that we use, the function is linear and the vector of friendliness parameters of P other funds is summarized by their average, n i, j kaj nk .
and then form the estimate of the unobserved component of director quality as ^ z i 1X ^ r : o o i, k ni kaj i, k 22
^ are in order. Three comments about our estimate z i First, since zi is only dened up to a re-scaling factor, we need to pick a unit of measurement for it. We do so by explicitly setting g3 1. Second, in our data, we have a median of 107 funds voting on a director, and so our estimates of unobserved director quality are quite precise: P they differ from actual qualities by 1=n kaj Zik , which becomes small as n becomes large. Finally, the averaging of the residuals is a departure from the standard control function approach. We average the residuals because the unobserved director quality is not specic to every observation; rather, it is specic to a director.18 6.2. Second stage We now get back to estimating the parameters of a funds best response function, ( ) 1 if a b1 nj b2 o e i, j b3 zi Gxi eij Z 0 oij : 23 0 if a b1 nj b2 o e i, j b3 zi Gxi eij o 0
oe i, j d g1 nij g2 n i, j G1 xi g3 zi :
14
17 Generally, o e i, j could also depend on which equilibrium is played by the funds, in the case of multiple equilibria. We do not investigate the issue of equilibrium selection here; instead, we assume that the same equilibrium is played in all elections, e.g., the most or the least management-friendly one.
18
Cooley (2010) uses a similar modication of the rst stage.
105
Note that since we set the scale for zi by setting g3 1, r we have o e i, j o i, j zi . Hence, we can rewrite the above equation as ( ) 1 if a b1 nj b2 o r i, j b2 b3 zi Gxi eij Z 0 o : ij 0 if a b1 nj b2 o r i, j b2 b3 zi Gxi eij o 0 24 In the last step of our estimation procedure, we replace variables o r i, j and zi in Eq. (24) with their rst-stage ^ and run a logistic regression implied ^ r and z estimates o i i, j by the resulting equation (votes o ij on the left-hand side ^ , and x on the right-hand side) to ^r ,z and variables nj , o i i i, j estimate coefcients a, b1 , b2 , b3 , and G in that equation. Note that Eq. (24) makes explicit the two channels through which unobserved director quality impacts the vote of fund i: directly, by changing its utility from voting for a director (measured by coefcient b3 ) and indirectly, by changing the expected votes of other funds (measured by coefcient b2 ), which in turn impacts fund is incentives. Like the rst stage, the second stage of our estimation procedure also departs from the standard control function approach: we use the expected votes of other funds rather than the actual votes. This is a consequence of modeling voting as a game of private information, in which each fund chooses its action based on the expectation of other funds behavior rather than the actual realization, which is unknown to the fund when it makes its decision. 7. Results We can now estimate the model of voting presented in Section 5. The object we are estimating is the funds best response function. In other words, we are estimating how, in equilibrium, a funds vote changes in response to changes in the expectation of the other funds votes, the type of director that is being voted on, and the funds own management-friendliness. 7.1. First stage In the basic specication of the rst stage, we regress the average vote of other funds in the election, o i, j , on a funds own management-friendliness, nij , other funds average management-friendliness, n i, j , and rm and director controls, xi (Eq. (19)).19 The results are reported in columns 1 (without rm and director controls) and 2 (with controls) of Table 7. In the alternative specication, we also include three higher moments of the distribution of other funds friendliness. The results for this specication are in columns 3 and 4. The average past voting record of funds voting on the director is strongly correlated with their vote both when included alone and with other moments of the distribution. For example, in the basic specication with rm and director controls (column 2), the coefcient is 0.77, and it
19 Recall from Section 4.1 that we compute a funds managementfriendliness as the average vote of that fund in rms other than the rm under observation in the previous year.
increases to 1.32 when higher moments are included (column 4). It is highly statistically signicant at 1% across all specications. From columns 3 and 4 we can see that the standard deviation, skewness, and kurtosis of other funds friendliness all contain information about the equilibrium vote of other funds, which is not contained in the mean. To uncover the unobserved director quality, we take the residual from the regression above and average it over the funds voting on a director, as in Eq. (22). A high residual from the rst stage tells us that the director obtained a higher vote than predicted by her characteristics and the characteristics of the funds which voted on her. 7.2. Second stage: estimating the parameters of the model We now have all determinants of a funds vote from our model. Following Eq. (24), we estimate it using a logistic regression where the dependent variable is the fund vote. The explanatory variables are the funds own friendliness, the expected average vote of the other funds, the estimate of director unobserved quality, and rm and director controls. 7.2.1. Strategic complementarity The expected average vote of other funds is a strong determinant of fund voting: the strategic effect is highly statistically signicant across specications. Results are presented in Table 8. Once we control for rm or director characteristics in these specications, the coefcient is 1.84 in the specication using only the mean friendliness of other funds as an instrument and 3.54 for the specication using the moments of other funds friendliness as instruments. The expectation of the average vote of other funds is determined in equilibrium and is therefore endogenous to the model. To shed some light on the economic magnitude of this coefcient, one has to consider an out-ofequilibrium shock to the expectation of other funds average vote. Consider the following experiment: how would a funds vote change if it made a mistake and overestimated the expected average vote of other funds. First, consider the higher coefcient, 3.54. If a funds expectation of the average vote of other funds increases by 10%, the log odds of a for vote increase by 0.354. Consider a fund which withheld its vote on a director with a 20% probability. Suppose that in equilibrium the fund expected all other funds to withhold their votes with a 20% probability as well. If this fund instead made a mistake and expected all other funds to withhold their votes with only a 10% probability, its own probability of voting withhold would decrease by 5.1% to 14.9%. Alternatively, if the coefcient is 1.84, then the same 10% change in the expectation of other funds average vote would cause a fund to change its probability of withholding by 2.8%, to 17.2%. An alternative way of interpreting the coefcients is a back-of-the-envelope calculation of the corresponding social multiplier. Suppose the direct effect of changing a
106
Table 7 First stage: equilibrium projection. The dependent variable is the average vote of other funds voting on the director in the shareholder meeting, where the for vote takes the value of one and zero otherwise. Moments of other funds management-friendliness are the moments of management-friendliness of funds voting on the director in the shareholder meeting. Own management-friendliness is calculated as the average vote of a fund in the previous calendar year on rms other than the rm under observation. Data sources: SEC Edgar (N-PX lings), CRSP, Compustat, IRRC governance database, Board Analyst directors database. (1) Other funds avg. vote for Moments of other funds management-friendliness Average Standard deviation Skewness Kurtosis Own management-friendliness Firm and director characteristics Constant Observations R2 0.002 [0.001] 0.047 [0.060] 1,766,982 0.02 0 [0.001] Y 0.249 [0.091] 1,493,621 0.09 (2) Other funds avg. vote for (3) Other funds avg. vote for (4) Other funds avg. vote for
1.067 [0.066]
0.766 [0.101]
1.426 [0.156] 0.431 [0.102] 0.005 [0.006] 0.001 [0.001] 0.005 [0.001] 0.431 [0.145] 1,766,982 0.03
1.32 [0.203] 0.489 [0.129] 0.011 [0.006] 0.001 [0.001] 0.004 [0.001] Y 0.292 [0.188] 1,493,621 0.09
Robust standard errors clustered by director election in brackets.
director characteristic decreases all funds probability of voting withhold from 20% to 10%. In addition to that response, each fund now knows that all other funds will lower their vote by 10%. If the coefcient is 3.54, each funds best response is to decrease its log odds of a withhold vote further by 3.5410% = 0.354, shifting the probability of a withhold vote from 10% to 7.2%. But then all funds know that as well, so they have an incentive to reduce their vote even further: an additional iteration decreases the log odds by (107.2%)3.54 = 0.10, decreasing the probability of funds voting withhold from 7.2% to 6.60%. Due to the structure of the game, by repeating the iterations until convergence, we can compute a new equilibrium. In this equilibrium, each fund votes withhold with probability 6.43%. Thus, the 10% direct effect of changing director quality translates to a total effect of 13.6%: the social multiplier is 1.36, and the complete effect of changing a director characteristic is more than a third larger than it would be in the absence of funds strategic interactions. Alternatively, with the estimated coefcient of strategic interaction equal to 1.84, the corresponding social multiplier would be equal to 1.18. In Section 8, we provide additional insight into the economic signicance of our estimates of strategic complementarity by analyzing its impact on equilibrium behavior of the funds.
have to jointly estimate all factors that determine a funds vote. Funds friendliness remains a statistically and economically signicant determinant of voting in the structural model. From Table 8, we can see that the coefcient on a funds own friendliness is remarkably stable across different specications of controls, ranging from 5.81 to 5.96. To interpret this effect, consider changing a funds friendliness by two standard deviations of the fund friendliness distribution. This change increases a funds log odds of a for vote by 1.86. To put this magnitude in perspective, suppose a fund at one-standard-deviation under the mean of friendliness were to withhold the vote with a 20% probability. If the fund were to change to a one-standard-deviation above the mean friendly fund, it would withhold with only 3.8% probability. Given the relevant range of voting outcomes, such a change in fund behavior is potentially very large. Furthermore, we have only considered the change in voting of a single fund, ignoring the equilibrium reinforcement that would take place were all funds to become more managementfriendly at the same time. As we will see below, heterogeneity in managementfriendliness is at least as important in determining fund voting as rm performance and director and governance characteristics. 7.2.3. Firm and director characteristics A change in rm or director characteristics has a direct and indirect effect. The coefcients of the best response function which we interpret in this section represent the magnitudes of the direct effect, which gives us a lower bound on the equilibrium size of the effect. Furthermore, all direct effects are reinforced by the same mechanism of
7.2.2. Fund heterogeneity In Section 4.1, we showed that a funds voting record from the previous year is a good proxy for how friendly a fund is, and that this friendliness affects fund voting. To understand the magnitude of this effect, however, we
107
Table 8 Best response function. The dependent variable is a vote cast by a fund in a board of directors election; it takes the value of one if the vote is for and zero otherwise. The predicted vote of other funds is the tted value of the corresponding column in Table 7. Unobserved quality is the average residual from the specication of the corresponding column in Table 7, averaged within a director-meeting pair. Own management-friendliness is calculated as the average vote of a fund in the previous calendar year on rms other than the rm under observation. The industry return is the value-weighted return of the rms two-digit SIC industry. The governance index is the governance index from Gompers, Ishii, and Metrick (2003). The omitted category from Inside director and Outside related director is an Outside director. Data sources: SEC Edgar (N-PX lings), CRSP, Compustat, IRRC governance database, Board Analyst directors database. (1) Vote for Predicted vote of other funds Own management-friendliness Unobserved quality Firm characteristics ROA Last year return Industry return Return below industry quartile Log assets Q Book-to-market Leverage Cash ow-to-assets Capex-to-assets S&P 500 Governance index Director characteristics Inside director Outside related director CEO Chairman Founder Audit member Compensation chair Compensation member Executive chair Executive member Governance chair Governance member Nominating chair Nominating member 5.608 [0.310] 5.834 [0.041] 7.824 [0.032] (2) Vote for 1.839 [0.650] 5.958 [0.047] 7.85 [0.038] 3.207 [0.423] 0.077 [0.020] 0.077 [0.045] 0.058 [0.024] 0.021 [0.009] 0.026 [0.009] 0.123 [0.034] 0.333 [0.055] 1.753 [0.420] 1.217 [0.249] 0.212 [0.034] 0.011 [0.004] 0.273 [0.053] 0.653 [0.075] 0.157 [0.049] 0.036 [0.039] 0.066 [0.060] 0.107 [0.021] 0.02 [0.050] 0.405 [0.029] 0.043 [0.093] 0.071 [0.026] 0.23 [0.119] 0.004 [0.034] 0.314 [0.113] 0.084 [0.035] (3) Vote for 6.349 [0.292] 5.813 [0.041] 7.817 [0.032] (4) Vote for 3.541 [0.539] 5.942 [0.047] 7.849 [0.038] 2.867 [0.413] 0.066 [0.019] 0.055 [0.045] 0.064 [0.024] 0.019 [0.009] 0.022 [0.009] 0.11 [0.034] 0.27 [0.053] 1.544 [0.414] 1.094 [0.245] 0.153 [0.032] 0.009 [0.003] 0.177 [0.049] 0.468 [0.063] 0.116 [0.048] 0.008 [0.039] 0.03 [0.058] 0.083 [0.021] 0.03 [0.050] 0.353 [0.027] 0.067 [0.092] 0.055 [0.026] 0.201 [0.118] 0.011 [0.034] 0.288 [0.112] 0.054 [0.034]
108
Table 8 (continued ) (1) Vote for Constant Observations 7.006 [0.282] 1,766,982 (2) Vote for 3.922 [0.587] 1,493,621 (3) Vote for 7.667 [0.268] 1,766,982 (4) Vote for 5.436 [0.496] 1,493,621
Robust standard errors clustered by director election in brackets. Signicant at 10%; signicant at 5%; signicant at 1%.
strategic voting. Therefore, if the direct effect of a change in last years return is economically greater than the direct effect of changing governance characteristics, then the overall effect of increasing last year returns on voting will also be larger than the overall effect of changing governance characteristics on voting. The rst group of effects we look into are measures of the rms last year performance. The previous year return on assets seems to be the most robust performance measure that funds respond to: the coefcient of 2.867 in specication 4 (Table 8) implies that a two-standarddeviation move in the ROA increases the log odds ratio of a director obtaining a for vote by 0.58. Returning to our canonical example, this change in returns would decrease the likelihood of a withhold vote of a fund from 20% to 12.3%. We also examine whether funds consider rm performance relative to the industry when casting votes. If a rm manages to switch from the bottom return quartile of its two-digit SIC industry, the log odds of a withhold vote decrease by 0.06. In our example, this would reduce the probability of a withhold vote from 20% to 19.0%. Directors of larger rms and rms that are members of the S&P 500 also obtain more favorable voting outcomes. There are several potential explanations for this effect: rst, there is a larger free-rider problem among the shareholders in large rms, which decreases the benet from voting withhold. Second, business connections with larger rms may be more important for funds. Third, the management of large rms might also have more funds at their disposal to hire proxy solicitors and investor relations rms to help them manage the voting. A coefcient of 0.153 implies that membership in S&P 500 would reduce a directors probability of a withhold from 20% to 17.7% in our example. Similarly, the coefcient on log assets of 0.019 implies that a two-standard-deviation increase in rm assets decreases a directors probability of a withhold vote by 2.9% in our canonical example. Next, we examine the effect of governance characteristics. Two outcomes are possible: funds could withhold their vote more frequently in rms with weaker shareholder rights to force the directors to increase shareholder rights. Alternatively, funds may withhold their vote more frequently in rms with stronger shareholder rights, where they may be able to exert more pressure on directors. The latter seems to be the case in the data. Funds are more likely to vote for a director in a more dictatorial rm, as measured by the GompersIshii-Metrick (GIM) index (Gompers, Ishii, and Metrick, 2003). A two-standard-deviation (5-point) increase in the index increases the log odds of a for vote by 0.05.
The direct effect is economically small: a two-standarddeviation increase in the GIM index decreases a funds probability of a withhold vote from 20% to 19.3%. A similar phenomenon, where funds seem to vote against directors on whom they may be able to exert more pressure, can also be seen in voting on the CEO. The CEO, on average, realizes fewer withhold votes than other inside directors. If we compare a CEO to a similar inside director whose probability of a withhold vote is 20%, the CEO is 1.8% less likely to receive a withhold vote. Funds frequently address a directors relationship with the rm as a reason to withhold the vote in their guidelines. Funds are least likely to withhold their vote on an outside director. Outside related directors, however, are more likely to obtain a withhold vote than inside directors. A potential explanation is that mutual funds understand that both outside related and inside directors are not free of conicts of interests. Outside related directors have available substitutes in outside directors. The latter may not be good substitutes for inside directors, who have signicantly more information on the company. An outside directors log odds ratio of a for vote is 0.18 higher than that of an inside director in the same rm. To put this coefcient in context, consider an inside director that receives a 20% withhold vote; the same director as an outsider would obtain a 17.3% withhold vote, leading to a 13% decrease in the probability of a withhold vote. Funds proxy guidelines also target committee members and specify that funds are to withhold votes if they are dissatised with the decisions or the composition of the committee. Compensation committee members receive, on average, the lowest withhold votes among committee members; a director whose withhold vote would otherwise be 20% receives 6.2% fewer votes if she is on the compensation committee. Similar in magnitude is the effect of being the chair of the nominating committee. We would like to caution that these results may be sensitive to the time period under consideration, due to the funds changing focus on different governance characteristics over time. For example, an accounting scandal may lead to more attention to the members of audit committees, making them subject to more withhold votes in the future. A longer time series may shed more light on the robustness of these effects. 7.2.4. Unobserved director quality Unobserved director quality plays an important role in determining fund voting. Note that the coefcient on unobserved director quality obtained from the logistic regression (Table 8) is not the structural coefcient.
109
To compute the latter, we need to adjust the regression coefcient by subtracting from it the coefcient on expected vote of other funds (see Section 6 and Eq. (24) for details). The magnitude of the structural coefcient is economically important. For example, consider the estimates from specication 4. These estimates imply that a two-standard-deviation change in unobserved director quality increases log odds of a for vote by 1.36, which is twice as large as the change following a two-standard-deviation increase in the return on assets. This result demonstrates the importance of accounting for unobserved director heterogeneity in our estimation of strategic complementarities, which would otherwise be severely biased.
, t 1 iteration t 1, o e i, j . Formally, , t 1 t a b1 nj b2 o e cij i, j b3 zi Gxi , t 1 X 1 X expcik Prot ik 1 t : ni 1 kaj ni 1 kaj 1 expcik
25 26
,t oe i, j
We start by setting the belief of each fund about the ,0 strategies of other funds to o e and iterate the i , j 1 P t following two equations until i, j jProij 1 1 1 j is less than 0 : 001. We repeat the procedure Prot ij ,0 by starting with o e i, j 0 to obtain the lowest equilibrium. 8.2. Performance of the model We begin by simulating equilibrium voting behavior under the estimated parameters, and compare it to the actual votes. The highest and lowest simulated equilibria are very close, and so we present the results only for the highest equilibrium. We calculate director support by averaging the probability of a withhold vote for all funds voting on a director. Panel A of Table 9 describes the distribution of the percentage of funds supporting a director in an election for the actual data and the simulated equilibrium under two specications. The model matches the qualitative features of the data. It also matches the quantitative features of vote distribution for directors who received support above the 10th percentile. The median director obtains withhold votes from 3.4% of funds in the actual data and 3.6% and 3.8% of funds in the simulated equilibria. The biggest difference between the realized vote and the predicted vote is in the level of support obtained by the 10% of directors with the highest level of opposition. In the data, the director at the 10th percentile of the distribution receives 26.4% withhold votes, while in our simulated equilibria he would receive such votes from 15.3% and 16.5% of funds. For the directors even further out in the tail, our models quantitative performance deteriorates further, most likely as a consequence of the structural assumptions in our estimation, e.g., the lack of heterogeneity in strategic complementarities among funds. We have to keep this caveat in mind when interpreting the counterfactual results. Our model does not replicate the empirical fact that many directors obtain no withhold votes. The reason is that we are comparing an actual realization of votes to the average probability of funds supporting a director. Using a logit model, no fund votes for a director with probability one; therefore, the mean expected support is lower than one by construction. 8.3. Social multiplier Suppose the quality of a director increases for an exogenous reason. Then a fund will have an incentive to increase its vote on the director simply because she is a better director. This is the direct effect of increasing quality. Furthermore, the fund will have an additional incentive to increase its vote, because it knows that other
8. Equilibrium impact of strategic complementarity We now use the estimated model to generate counterfactuals to examine the equilibrium consequences that changes in underlying parameters have on the voting outcome. Factors that directly affect fund voting are amplied in equilibrium. We briey explored the magnitude of this multiplier in Section 7. In this section, we analyze it in more detail.
8.1. Constructing counterfactuals To construct counterfactuals, we alter the inputs used in the model we estimated and nd a Bayesian Nash equilibrium of the resulting game. Because of the supermodular structure of the game, an equilibrium exists, and there is a simple algorithm that nds the most and the least management-friendly equilibria. We rst specify funds strategies and beliefs. Using Remark 1, we can specify a strategy of fund j, given its own type, director characteristics, and other funds expected average vote, by a cutoff cij. For realizations of the idiosyncratic shock eij below cij, the fund votes withhold; otherwise it votes for. The cutoff for fund j can be obtained from our best response function (13) as cij a b1 nj b2 o e i, j b3 zi Gxi . The logistic specication of the best response function allows us to obtain a closed-form expression for the expected vote corresponding to the funds strategy Proij 1 expcij =1 expcij . Moreover, from (13) we know that the sufcient statistic for a funds belief about other funds strategies is the expected average vote of other funds o e i , j P P 1=ni 1 kaj Proik 1 1=ni 1 kaj expcik =1 expcik . The highest equilibrium is obtained by iterating the best response function from the highest strategy prole. t be the cutoff value for the strategy and Pr ot Let cij ij 1 the expected vote of fund j after the t th iteration of the best response function. In t th iteration we calculate t . the strategy of every fund by calculating its cutoff, cij The belief of each fund about other funds strategies is the strategy of the other funds from the previous iteration, i.e., it is the expected vote of other funds xed from
110 Table 9 Simulations. Panel A compares the distribution of director votes in the sample to the distribution of director votes in the simulated equilibria based on parameter estimates in specications 2 and 4 from Table 8. Director actual vote is the average withhold vote obtained by directors. Director simulated vote is the average probability of a withhold vote for a director computed using her actual attributes and estimated parameters from specications 2 and 4 in Table 8, respectively. The sample is restricted to directors for whom we were able to obtain counterfactuals in all specications. Panel B describes simulated equilibria for two different specications of parameters on a subset of directors who were in the lowest quartile of unobservable director quality in that specication. The parameters for the simulation are taken from specications 2 and 4 in Table 8. Director simulated vote is the average withhold vote for a director computed using her actual attributes. Direct impact vote is the director simulated vote increased by the odds difference implied by an increase in unobservable quality of 9010 percentile range in unobservable quality. Equilibrium impact vote is the vote computed from our model using the directors actual attributes, but increasing her unobserved quality by the 9010 percentile range in unobservable quality. Panel C presents the distribution of the multiplier for each specication, which is calculated as (Equilibrium impact voteDirector simulated vote)/(Direct impact voteDirector simulated vote). All simulated equilibria presented are the most management-friendly equilibria. Variable N Mean St. dev. p10 p25 p50 p75 p90 G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112
Panel A: Actual and simulated distributions of votes Director actual vote 7,897 Director simulated vote (Spec. 2) 7,897 Director simulated vote (Spec. 4) 7,897 Panel B: Simulated equilibria for low-quality directors Specication 2 Director simulated vote 2,300 Direct impact vote 2,300 Equilibrium impact vote 2,300 Specication 4 Director simulated vote 2,302 Direct impact vote 2,302 Equilibrium impact vote 2,302 Panel C: Social multiplier for low-quality directors Multiplier (Spec. 2) 2,300 Multiplier (Spec. 4) 2,302
9.5% 9.9% 11.0%
16.7% 18.9% 22.7%
26.4% 16.5% 15.3%
7.9% 5.5% 5.3%
3.4% 3.8% 3.6%
1.4% 2.8% 2.7%
0.0% 2.2% 2.1%
28.6% 13.3% 8.3% 33.9% 22.4% 8.8% 1.16 1.64
31.0% 18.6% 12.1% 37.5% 29.6% 14.7% 0.21 0.92
80.9% 43.9% 24.0% 92.2% 72.7% 22.7% 1.01 1.02
56.6% 21.3% 11.7% 79.9% 48.8% 10.4% 1.02 1.04
9.1% 2.5% 2.2% 8.7% 2.9% 2.5% 1.03 1.07
4.4% 1.0% 0.9% 4.3% 1.3% 1.1% 1.26 2.13
2.9% 0.6% 0.6% 2.9% 0.8% 0.7% 1.53 3.33
111
funds are also more likely to vote for the director, amplifying the direct effect. In this section, we compare the direct effect to the full equilibrium effect. Our approach allows us to obtain director-specic multipliers, which incorporate director characteristics and the distribution of fund friendliness. We take directors in the lowest quartile of unobserved quality. For each director, we increase her quality by the difference in unobserved quality between the 90th- and the 10th-percentile directors. We compare the direct effect of this change to the full effect. As a baseline, we use the votes obtained from the simulated equilibria described in the previous section. We use the simulated rather than the actual votes as a baseline, because we want to isolate the effects of the multipliers, and do not want to confound the effects with how well our model can replicate the voting distribution. Panel B of Table 9 presents the distributions of simulated votes taking into account the direct effect and the full effect of increasing director quality.20 The full effect is much stronger than the direct effect. For example, in specication 4, the mean withhold vote drops from 33.9% in the baseline to 22.4% due to the direct effect and to 8.8% under full equilibrium response. This implies an aggregate multiplier of 2.2. The distributions of director-specic multipliers are presented in Panel C. The distributions are highly skewed. The median multipliers in our specications are 1.03 and 1.07. The multipliers in the 75th percentiles are much higher, at 1.26 and 2.13. In an earlier version of this paper, we reported additional results using various more exible specications of the model, suggesting that the higher estimate is more robust. Hence, for many directors (at least those above the 75th percentile), the social multiplier is likely to be largeat least two. Thus, any analysis of policy changes or improvements in rm or director characteristics has to take into account the fact that peer effects have an important inuence on voting outcomes and the resulting social multipliers differ widely across directors. 9. Conclusion Voting in director elections is an important tool that shareholders have at their disposal. Prior empirical literature on boards of directors has used outcome variables such as board composition and director survival rates to understand this process. We explore shareholder voting using a novel, comprehensive data set which includes 2,058,788 mutual fund votes in director elections that took place between July 1, 2003 and June 30, 2005. Looking at individual votes allows us to study the behavior of the funds and the interactions among them. We nd substantial systematic heterogeneity in voting patterns, with some funds being consistently more
20 Formally, if oe ij is the initial probability of fund i voting for ^ is the coefcient on director j, Dz is the change in director quality, and b 3 unobserved quality from (13), then the new probability of voting for e ^ resulting from the direct effect is explnoe ij =1oij b 3 e ^ Dz: Dz=1 explnoe = 1 o b 3 ij ij
management-friendly than others. This heterogeneity has a large effect on voting, on par with rm and director characteristics. Thus, who monitors directors can potentially be as important as the characteristics of the directors monitored. We then estimate a model of voting in which mutual funds impose externalities on each other: the cost of opposing management decreases when other funds oppose it as well. To estimate the model, we exploit fund heterogeneity. Using this variation, we show that strategic interaction between funds is an economically and statistically signicant factor in determining fund voting. It reinforces the direct effect that director characteristics have on the voting outcome. This social multiplier varies substantially across directors. The large and persistent differences in voting patterns and the strategic interaction effects suggest that at least some funds take voting seriously and try to use it to inuence management. Yet on its own, voting in director elections is essentially powerless, since under the plurality voting system implemented in most U.S. companies, it takes just one for vote to elect a director. The reason why votes nevertheless have power is that shareholders have other tools at their disposal that they can use to inuence managers and directors, if voting outcomes are ignored. Investors can sell shares in the company. They also can, at considerable cost, propose an alternative slate of directors and launch a proxy ght. A clear expression of dissatisfaction with the management coming from the shareholders in the form of low director support could also send a signal to outsiders that the company is an attractive takeover target (dissatised shareholders will be more willing to agree to a takeover offer and sell their shares instead of sticking with the current management). These examples suggest that votes play an important role as signals, and empirically studying this role and the interaction of voting and other tools of shareholder control is an important direction for future research. As more data on mutual fund voting become available,21 it becomes feasible to study the dynamic interaction between shareholder voting behavior, management response to the voting outcomes, and the use of other tools of corporate governance. Better understanding of the role of voting as part of the larger corporate governance mechanism will in turn make it possible to address important policy questions. For instance, instead of the plurality system, some U.S. companies have begun to adopt majority voting systems. Under majority voting, withhold votes have the weight of an against vote and a candidate gets elected only if more than half of the votes cast are for. A number of shareholder advocates and policymakers have recently called for a broader use of the majority voting system by corporations or even a full elimination of plurality voting (ISS, 2005). Viewing voting in isolation, majority voting seems to give shareholders much more power. In the richer framework, however, the question of which
21 As of August 31, 2009, six years of voting data are available from the SEC.
112
mechanism allows shareholders to exert more inuence on directors and managers becomes more subtle. For moderate levels of dissatisfaction with directors, if shareholders simply want to signal that they would like the director or the management to improve their performance but do not want to re the director, they may be willing to withhold support under the plurality system but not under the majority system, making the latter less effective. Therefore, determining which of the two mechanisms is more effective becomes an empirical question, requiring careful analysis and better understanding of the role of voting in director elections as part of the broader set of tools of shareholder control. References
Ackerberg, D., Benkard, C., Berry, S., Pakes, A., 2007. Econometric tools for analyzing market outcomes. In: Heckman, J., Leamer, E. (Eds.), Handbook of Econometrics, vol. 6A. North-Holland, Amsterdam, pp. 41714276 (Chapter 63). Cai, J., Garner, J., Walkling, R., 2009. Electing directors. Journal of Finance 64, 23892421. Cooley, J., 2010. Desegregation and the achievement gap: Do diverse peers help? Unpublished working paper. University of Wisconsin, Madison. Davis, G., Kim, E., 2007. Business ties and proxy voting by mutual funds. Journal of Financial Economics 85, 552570. Feddersen, T., Pesendorfer, W., 1997. Voting behavior and information aggregation in elections with private information. Econometrica 65, 10291058. Fischer, P., Gramlich, J., Miller, B., White, H., 2009. Investor perceptions of board performance: evidence from uncontested director elections. Journal of Accounting and Economics 48, 172189. Glaeser, E., Sacerdote, B., Scheinkman, J., 1996. Crime and social interactions. Quarterly Journal of Economics 111, 507548. Goetzmann, W., Brown, S., 1997. Mutual fund styles. Journal of Financial Economics 43, 373399. Gompers, P., Ishii, J., Metrick, A., 2003. Corporate governance and equity prices. Quarterly Journal of Economics 118, 107155.
Heard, J., Sherman, H., 1987. Conicts of Interest in the Proxy Voting System. Investor Responsibility Research Center, Washington, DC. Hermalin, B., Weisbach, M., 2003. Boards of directors as an endogenously determined institution: a survey of the economic literature. Economic Policy Review 9, 726. Hong, H., Kubik, J., Stein, J., 2004. Social interactions and stock-market participation. Journal of Finance 59, 137163. Hong, H., Kubik, J., Stein, J., 2005. Thy neighbors portfolio: word-ofmouth effects in the holdings and trades of money managers. Journal of Finance 60, 28012824. Imbens, G., Wooldridge, J., 2007. Control Function and Related Methods. NBER Summer Institute Lecture Notes. Available at /http://www. nber.org/WNE/lect_6_controlfuncs.pdfS. Institutional Shareholder Services, 2005. Majority voting in director elections: from the symbolic to the democratic. The ISS Institute for Corporate Governance White Paper. Malmendier, U., Tate, G., 2008. Who makes acquisitions? CEO overcondence and the markets reaction. Journal of Financial Economics 89, 2043. Matvos, G., Ostrovsky, M., 2008. Cross-ownership, returns, and voting in mergers. Journal of Financial Economics 89, 391403. McGurn, P., 1989. Condential Proxy Voting. Investor Responsibility Research Center, Washington, DC. Monks, R., Minow, N., 2003. Corporate Governance, third ed. Blackwell Publishers, Malden, MA. Navarro, S., 2008. Control functions. In: Durlauf, S., Blume, L. (Eds.), New Palgrave Dictionary of Economics, second ed. Palgrave Macmillan, London. NYSE, 2006. NYSE adopts proxy working group recommendation to eliminate broker voting in 2008. NYSE press release. Available at /http://www.nyse.com/press/1161166307645.htmlS. Rothberg, B., Lilien, S., 2006. Mutual funds and proxy voting: new evidence on corporate governance. Journal of Business and Technology Law 1, 157184. Sacerdote, B., 2001. Peer effects with random assignment: results for Dartmouth roommates. Quarterly Journal of Economics 116, 681704. Securities and Exchange Commission, 2003. Proxy voting by investment advisers. SEC Release No. IA-2106; File No. S7-38-02. Available at /http://www.sec.gov/rules/nal/ia-2106.htmS. Van Zandt, T., Vives, X., 2007. Monotone equilibria in Bayesian games of strategic complementarities. Journal of Economic Theory 134, 339360.

Matvos, Ostrovsky (2010) Mutual Fund Proxy Voting

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Matvos, Ostrovsky (2010) Mutual Fund Proxy Voting

Uploaded by

Copyright:

Available Formats

Journal of Financial Economics 98 (2010) 90112

Contents lists available at ScienceDirect

Journal of Financial Economics

Heterogeneity and peer effects in mutual fund proxy voting$

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

2.53% (***) 3.71% (***) 0.37% (***)

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

0.429 [0.0039] 1,766,982 0.11 Y

Signicant at 10%; signicant at 5%; signicant at 1%.

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

Signicant at 10%; signicant at 5%; signicant at 1%.

Robust standard errors clustered on fund-year in brackets.

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

each fund realizes its payoff U qi , nj , oij , oi, j , eij : 1

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

 Condition 1 requires supermodularity and increasing

Third, we normalize a funds utility from voting withhold to zero: U xi , zi , nj , 0, o i, j , eij 0: 9

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

We can now rewrite the actual average vote of other funds as

d g1 nij g2 n i, j G1 xi g3 zi Zij d g1 nij g2 n i, j G1 xi g3 zi Zij or i, j rij ,

Cooley (2010) uses a similar modication of the rst stage.

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

Signicant at 10%; signicant at 5%; signicant at 1%.

Robust standard errors clustered by director election in brackets.

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

, t 1 iteration t 1, o e i, j . Formally, , t 1 t a b1 nj b2 o e cij i, j b3 zi Gxi , t 1 X 1 X expcik Prot ik 1 t : ni 1 kaj ni 1 kaj 1 expcik

9.5% 9.9% 11.0%

16.7% 18.9% 22.7%

26.4% 16.5% 15.3%

7.9% 5.5% 5.3%

3.4% 3.8% 3.6%

1.4% 2.8% 2.7%

0.0% 2.2% 2.1%

28.6% 13.3% 8.3% 33.9% 22.4% 8.8% 1.16 1.64

31.0% 18.6% 12.1% 37.5% 29.6% 14.7% 0.21 0.92

80.9% 43.9% 24.0% 92.2% 72.7% 22.7% 1.01 1.02

56.6% 21.3% 11.7% 79.9% 48.8% 10.4% 1.02 1.04

9.1% 2.5% 2.2% 8.7% 2.9% 2.5% 1.03 1.07

4.4% 1.0% 0.9% 4.3% 1.3% 1.1% 1.26 2.13

2.9% 0.6% 0.6% 2.9% 0.8% 0.7% 1.53 3.33

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

G. Matvos, M. Ostrovsky / Journal of Financial Economics 98 (2010) 90112

You might also like

2.53% () 3.71% () 0.37% (***)

Condition 1 requires supermodularity and increasing